Deep learning processing system based on Winograd

The invention belongs to the technical field of artificial intelligence processors, and particularly relates to a Winograd-based deep learning processing system, and the system comprises an input channel which communicates with external equipment; the input characteristic loader is connected with the input channel; the routing module is connected with the input layer loader; the convolution accelerator core module is connected with the routing module, and the convolution kernel accelerator core module comprises at least two convolution kernel accelerator sub-core modules which are arranged in parallel; the channel accumulator is connected with the convolution accelerator core module and is used for accumulating output data of the parallel convolution kernel accelerator sub-core modules; and the feature unloader is connected with the channel accumulator and transmits the output data of the channel accumulator to the outside of the processing system. Through a multi-size Wino-DPU hybrid architecture and optimal